Seedance 2.5 Pricing: Every API and Platform Compared (August 2026)

§ 01The model in one paragraph

Seedance 2.5 (ByteDance, launched July 31, 2026; API model dreamina-seedance-2-5-260628) generates 4 to 30 seconds of video at 24 fps with synchronized audio included, from up to 50 reference assets (30 images, 10 videos, 10 audio clips). The official API outputs 480p and 720p natively; that detail matters below, because the 1080p, 1440p and 4K tiers you will see on resellers are provider-side rollouts or super-resolution upscales, not documented native model output. For what the model can actually do, see our hands-on: full films from one prompt.

§ 02The full price table

Rates are per output second for jobs WITHOUT a video input, the cleanest comparison. Reference-video billing works differently and gets its own section below.

Three reading notes. First, the two BytePlus rows are different products: ModelArk is direct token-billed inference ($10.70 per million video tokens without video input), while LAS is a managed operator with pre/post-processing billed at $0.303 per second times resolution factors. Second, the ModelArk caveat: the launch rate is officially sourced, but the logged-out public ModelArk pricing table did not display a Seedance 2.5 row when we checked on August 22, so confirm the live number in your console before budgeting a production run. Third, WaveSpeed Turbo’s 1080p at $0.21/sec is the outlier worth testing yourself; verify whether it is native provider-side output or a separate serving path before you standardize on it.

§ 03The billing mechanics that decide your real bill

Everything is priced per second of output, and resolution multiplies it. A 720p second costs roughly 2.1 to 2.3 times a 480p second on every door. There is no per-clip flat fee anywhere in this table.

Audio is included. Seedance 2.5 generates synchronized speech, music and effects natively, and no provider publishes an audio surcharge.

Reference video changes the formula entirely. Give the model a video input (reference, edit target, or extension source) and nearly every provider switches to a LOWER per-second rate applied to input duration PLUS output duration:

The trap: a 10-second reference plus a 10-second output bills as ~20 counted seconds on most doors, so the “cheaper” reference rate frequently costs more per finished clip than plain text-to-video. Model your actual input lengths before picking a provider on the reference rate.

Promotions are dated. Atlas Cloud’s resolution menu carries -20% on native 1080p and -17% to -44% on its ESR upscale tiers, advertised through September 17, 2026 at 14:00 UTC+8; with the tags applied, its live calculator quoted us about $0.59/sec for native 1080p and about $1.70/sec for 4K ESR. EvoLink’s $0.528 1080p is likewise a 28% promotion with the same end date. Neither is a permanent base price.

§ 04Platform by platform: the API routes

BytePlus ModelArk / LAS (official, international). The source of truth for the model, first to get updates, and the launch token rate is the cheapest number in this article. Cons: not sold in the US, Canada, UK, Australia or New Zealand (a published country allowlist covers ~40 markets across Asia, the Middle East and Europe); the public pricing page had no visible 2.5 row when we checked; and the cheaper direct route may involve account verification. The LAS operator adds managed pre/post-processing at roughly double the direct rate.

Volcano Engine Ark (official, China). The mainland-China route (¥70/M tokens without video input, ¥42/M with). Requires real-name verification and RMB-oriented payment; overseas companies can apply via manual corporate review. Practical only if you already operate in China.

Replicate. The cheapest globally available door at the official launch rate, with a simple API and 63,000+ runs on the model page within three weeks. Cons: no published 1080p, and the steepest reference-video pricing in the table ($0.9676/sec at 720p), so keep it for text- and image-to-video jobs.

Atlas Cloud. We probed its live calculator directly on August 22, and the real model is per-resolution pricing applied identically across all three routes: about $0.14/sec at 480p and $0.30/sec at 720p, quoted the same on text-to-video and reference-to-video, billed on output seconds. It is also the only door advertising NATIVE 1080p generation (its docs are explicit that 1080p is a true Seedance render while its -sr/-esr tiers upscale from a lower native source), quoted at about $0.59/sec with a -20% promo tag, and its 4K ESR delivery quoted about $1.70/sec. Cons: the marketing cards say “from $0.134/sec,” which understates the calculator’s own quotes, so trust the number next to the Run button; no public numeric rate limits (handle 429s); and some requests route through third-party providers including BytePlus, which matters if you have data-residency requirements.

fal.ai. The most polished developer experience in the list and often the first reseller to ship new endpoints (1080p arrived around August 17). Cons: the most expensive mainstream 720p rate at $0.4730/sec, roughly double Replicate.

WaveSpeed. The widest endpoint menu (text, image, reference, extend, edit, plus Turbo variants) and the cheapest published 1080p-class route via Turbo. Cons: Standard-tier rates are mid-pack, and localized/cached pricing pages have shown inconsistent figures, so quote from the live calculator.

EvoLink, reAPI, Kie.ai. The budget tier. reAPI publishes the lowest reseller rates ($0.1186/$0.2668) but has a minimum billing floor on reference jobs. EvoLink has the broadest published resolution table including promotional 1080p. Kie.ai requires prepaid credits with a ~10% bulk bonus. All three are smaller operations with thinner public documentation than the platforms above; run a paid pilot before routing production volume.

§ 05The consumer platforms (and which ones have an API)

If you want a UI rather than an endpoint, seven platforms had confirmed Seedance 2.5 access on August 22:

Notable details: Higgsfield publishes the clearest consumer rates (3 credits/sec at 480p, 6.5 at 720p) but gates the model to its $49 Plus plan and up. Runway charges 20/30 credits per second (480p/720p) plus a surcharge per input-video second. Freepik is the only consumer platform with both a published per-second credit rate (200/440/1,100 at 480p/720p/1080p) and API access included on paid plans. OpenArt’s unlimited-2.5 promotion on Pro and Wonder plans ends August 23 at 11:59 p.m. Pacific, the day after this post publishes. And Dreamina, ByteDance’s own consumer app, has no public API at all; developer access lives on BytePlus, not in the app.

Genmo, Vadoo, InVideo and Canva showed no confirmed 2.5 access when we checked (InVideo’s docs name Seedance 2.0 only).

§ 06Who can buy where: country availability

• BytePlus (official international): published allowlist of roughly 40 markets including the EEA, Japan, Korea, Singapore, the Gulf states and Southeast Asia. NOT available in the United States, Canada, the UK, Australia or New Zealand.
• Volcano Engine (official China): real-name verification required, RMB payment; Chinese individuals and companies have automated paths, foreign enterprises need manual review.
• fal, Replicate, Atlas Cloud, WaveSpeed: available in the US and broadly worldwide, excluding US-sanctioned jurisdictions. Standard cards; no business verification for ordinary use.
• EvoLink, Kie.ai: apparently broadly self-service (EvoLink takes cards, Alipay and WeChat Pay), but neither publishes a country matrix; confirm at signup.

The practical consequence: US, UK, Canadian and Australian teams cannot buy the official door at the official price, which is exactly why the reseller market above exists and why reseller rates carry a margin over the ModelArk token rate.

§ 07Three worked scenarios

A 5-second 720p product clip, text-to-video. Replicate $1.16, reAPI $1.33, Atlas Cloud $1.50, Kie.ai $1.58, WaveSpeed $1.80 ($1.00 Turbo), BytePlus LAS $2.31, fal $2.37. On consumer platforms: about 150 Runway credits or 33 Higgsfield credits (~$1.60 on Plus).

A 30-second 720p narrative scene, the model’s signature trick. Replicate $6.94, reAPI $8.00, EvoLink $8.79, Atlas Cloud $9.03 (its calculator’s exact quote), WaveSpeed Turbo $6.00, BytePlus LAS $13.86, fal $14.19. This is where provider choice pays: the same clip varies by more than 2x.

A 10-second edit of a 10-second source video at 720p. Counted as ~20 seconds on most doors: EvoLink $3.58, Kie.ai $3.80, WaveSpeed $4.40, fal $5.68, BytePlus LAS $5.53. Atlas Cloud’s calculator quotes reference jobs at its output-second rates, which would put this at about $3.00 if input duration is not counted; verify on your own job before assuming it. Replicate’s $0.9676/sec makes it the wrong door for this job at ~$9.68.

§ 08The honest caveats

Every number here was verified against provider pages on August 22, 2026, and this market moves weekly: fal added 1080p in the week before publication, Atlas Cloud replaced its rate card with a promotional resolution table, and the official ModelArk page was mid-update when we checked. Two promotions expire September 17 and one August 23. Reseller “1080p” and “4K” tiers are provider-side additions to a model whose official spec is 480p/720p, and quality across those upscale paths is not equivalent. Treat the generation-time price shown by your provider as authoritative over any table, including this one, and re-verify before a large run.

If what you actually want is the finished film rather than the rate card, that is the job we built CellCog for: an agent that plans the shots, runs the generations, reviews the output and assembles the cut, priced by the work you assign rather than by the second of video.